Creating the Lustre Metadata Service (MDS)

From Lustre Wiki
Jump to: navigation, search

Syntax Overview

The syntax for creating an MDT is:

mkfs.lustre --mdt \
  [--reformat] \
  --fsname <name> \
  --index <n> \
  --mgsnode <MGS NID> [--mgsnode <MGS NID> …] \
  [ --servicenode <NID> [--servicenode <NID> …]] \
  [ --failnode <NID> [--failnode <NID> …]] \
  [ --backfstype=ldiskfs|zfs ] \
  [ --mkfsoptions <options> ] \
  <device path> | <pool name>/<dataset> <zpool specification>

The command line syntax for formatting an MDT incorporates several additional parameters when compared to that of the much simpler MGT. First, the metadata target requires a record of the NIDs that can provide the Lustre management service (MGS). The MGS NIDs are supplied using the --mgsnode flag. If there is more than one potential location for the MGS (i.e., it is part of a high availability failover cluster configuration), then the option is repeated for as many failover nodes as are configured (usually there are two).

Ordering of the MGS nodes in the command line is significant: the first --mgsnode flag must reference the NID of the current active or primary MGS server. If this is not the case, then the first time that the MDS tries to join the cluster, it will fail. The first time mount of a storage target does not currently check the failover locations when trying to establish a connection with the MGS. When adding new storage targets to Lustre, the MGS must be running on its primary NID.

The MDT must also be supplied with the name (--fsname) of the Lustre file system (maximum 8 characters), and a unique index number (--index) for the file system. There must always be an MDT with index=0 (zero) for each file system, representing the root of the file system tree. For many Lustre file systems, a single MDT (referred to as MDT0) is sufficient.

The list of service nodes (--servicenode) or failover nodes (--failnode) must be specified for any high availability configuration. Although there are more compact declarations for defining the nodes, for simplicity, list the NID of each server that can mount the storage as a separate --servicenode entry.

The next example uses the --servicenode syntax to create an MDT that can be run on two servers as an HA failover resource:

[root@rh7z-mds2 system]# mkfs.lustre --mdt \
>   --fsname demo \
>   --index 0 \
>   --mgsnode 192.168.227.11@tcp1 \
>   --mgsnode 192.168.227.12@tcp1 \
>   --servicenode 192.168.227.12@tcp1 \
>   --servicenode 192.168.227.11@tcp1 \
>   /dev/dm-2

The command line formats a new MDT that will be used by the MDS for storage. The MDT will provide metadata for a file system called demo, and has index number 0 (zero). There are two NIDs defined as the nodes able to host the MDS service, denoted by the --servicenode options, and two NIDs supplied for the MGS that the MDS will register with.

The --failnode syntax is similar, but is used to define only a failover target for the storage service. For example:

[root@rh7z-mds2 system]# mkfs.lustre --mdt \
>   --fsname demo \
>   --index 0 \
>   --mgsnode 192.168.227.11@tcp1 \
>   --mgsnode 192.168.227.12@tcp1 \
>   --failnode 192.168.227.11@tcp1 \
>   /dev/dm-2

Here, the failover host is identified as 192.168.227.11@tcp1, one server in an HA pair (and which, for the purpose of this example, has the hostname rh7z-mds1). The mkfs.lustre command was executed on rh7z-mds2 (NID: 192.168.227.12@tcp1), and the mount command must also be run from this host when the service starts for the very first time.

MDT Formatted as an LDISKFS OSD

The syntax for creating an LDISKFS-based MDT is:

mkfs.lustre --mdt \
  [--reformat] \
  --fsname <name> \
  --index <n> \
  --mgsnode <MGS NID> [--mgsnode <MGS NID> …] \
  [ --servicenode <NID> [--servicenode <NID> …]] \
  [ --failnode <NID> [--failnode <NID> …]] \
  [ --backfstype=ldiskfs ] \
  [ --mkfsoptions <options> ] \
  <device path>

The following example uses the --servicenode syntax to create an MDT that can be run on two servers as an HA failover resource:

[root@rh7z-mds2 system]# mkfs.lustre --mdt \
>   --fsname demo \
>   --index 0 \
>   --mgsnode 192.168.227.11@tcp1 \
>   --mgsnode 192.168.227.12@tcp1 \
>   --servicenode 192.168.227.12@tcp1 \
>   --servicenode 192.168.227.11@tcp1 \
>   --backfstype=ldiskfs \
>   /dev/dm-2

The --failnode syntax is similar, but is used to define only a failover target for the storage service. For example:

[root@rh7z-mds2 system]# mkfs.lustre --mdt \
>   --fsname demo --index 0 \
>   --mgsnode 192.168.227.11@tcp1 --mgsnode 192.168.227.12@tcp1 \
>   --failnode 192.168.227.11@tcp1 \
>   --backfstype=ldiskfs \
>   /dev/dm-2

The above examples are repeated from the main introduction to the syntax, but are included here to maintain symmetry with the rest of the text. However, note that the --backfstype flag has been set to ldiskfs, which tells mkfs.lustre to format the device as an LDISKFS OSD.

MDT Formatted as a ZFS OSD

Formatting an MDT using only the mkfs.lustre command

Note: For the greatest flexibility and control when creating ZFS-based Lustre storage targets, do not use this approach – instead, create the zpool separately from formatting the Lustre OSD. See Formatting an MDT using zpool and mkfs.lustre.

The syntax for creating a ZFS-based MDT using only the mkfs.lustre command is:

mkfs.lustre --mdt \
  [--reformat] \
  --fsname <name> \
  --index <n> \
  --mgsnode <MGS NID> [--mgsnode <MGS NID> …] \
  [ --servicenode <NID> [--servicenode <NID> …]] \
  [ --failnode <NID> [--failnode <NID> …]] \
  --backfstype=zfs \
  [ --mkfsoptions <options> ] \
  <pool name>/<dataset> \
  <zpool specification>

The next example uses the --servicenode syntax to create an MDT that can be run on two servers as an HA failover resource:

[root@rh7z-mds2 system]# mkfs.lustre --mdt \
>   --fsname demo \
>   --index 0 \
>   --mgsnode 192.168.227.11@tcp1 --mgsnode 192.168.227.12@tcp1 \
>   --servicenode 192.168.227.12@tcp1 \
>   --servicenode 192.168.227.11@tcp1 \
>   --backfstype=zfs \
>   demo-mdt0pool/mdt0 \
>   mirror sdb sdd

In addition to defining the parameters of the MDT service itself, the command defines a mirrored ZFS zpool called demo-mdt0pool consisting of two devices, and creates a ZFS dataset called mdt0. Normally, it is expected that the MDT will be created from a larger pool of storage, to maximize performance and meet requirements for capacity; the above example is provided for the purposes of outlining the command line syntax.

The --failnode syntax is similar, but is used to define only a failover target for the storage service. For example:

[root@rh7z-mds2 system]# mkfs.lustre --mdt \
>   --fsname demo --index 0 \
>   --mgsnode 192.168.227.11@tcp1 --mgsnode 192.168.227.12@tcp1 \
>   --failnode 192.168.227.11@tcp1 \
>   --backfstype=zfs \
>   demo-mdt0pool/mdt0 mirror sdb sdd

Note:: When creating a ZFS-based OSD using only the mkfs.lustre command, it is not possible to set or change some properties of the zpool or its vdevs, such as the multihost and ashift properties. For this reason, it is highly recommended that the zpools be created independently of the mkfs.lustre command, as shown in the next section.

Formatting an MDT using zpool and mkfs.lustre

To create a ZFS-based MDT, create a zpool to contain the MDT file system dataset, then use mkfs.lustre to create the actual file system dataset inside the zpool:

zpool create [-f] -O canmount=off -o multihost=on \
  [-o ashift=<n>] \
  -o cachefile=/etc/zfs/<zpool name>.spec | -o cachefile=none \
  <zpool name> \
  <zpool specification>

mkfs.lustre --mdt \
  [--reformat] \
  --fsname <name> \
  --index <n> \
  --mgsnode <MGS NID> [--mgsnode <MGS NID> …] \
  [ --servicenode <NID> [--servicenode <NID> …]] \
  [ --failnode <NID> [--failnode <NID> …]] \
  --backfstype=zfs \
  [ --mkfsoptions <options> ] \
  <pool name>/<dataset>

For example:

# Create the zpool
# Pool will comprise 3 mirrors each with 2 devices.
# Mirrors will be concatenated (striped).
zpool create -O canmount=off -o multihost=on \
  -o cachefile=none \
  demo-mdt0pool \
  mirror sdd sde mirror sdf sdg mirror sdh sdi

# Format MDT0 for Lustre file system "demo"
mkfs.lustre --mdt \
  --fsname demo \
  --index 0 \
  --mgsnode 192.168.227.11@tcp1 --mgsnode 192.168.227.12@tcp1 \
  --servicenode 192.168.227.12@tcp1 \
  --servicenode 192.168.227.11@tcp1 \
  --backfstype=zfs \
  demo-mdt0pool/mdt0

The output from the above example will look something like this:

# The zpool command will not return output if there are no errors
[root@rh7z-mds2 system]# zpool create -O canmount=off -o multihost=on \
>   -o cachefile=none \
>   demo-mdt0pool \
>   mirror sdd sde mirror sdf sdg mirror sdh sdi

[root@rh7z-mds2 system]# mkfs.lustre --mdt \
>   --fsname demo \
>   --index 0 \
>   --mgsnode 192.168.227.11@tcp1 --mgsnode 192.168.227.12@tcp1 \
>   --servicenode 192.168.227.12@tcp1 \
>   --servicenode 192.168.227.11@tcp1 \
>   --backfstype=zfs \
>   demo-mdt0pool/mdt0

   Permanent disk data:
Target:     demo:MDT0000
Index:      0
Lustre FS:  demo
Mount type: zfs
Flags:      0x1061
              (MDT first_time update no_primnode )
Persistent mount opts: 
Parameters:  mgsnode=192.168.227.11@tcp1:192.168.227.12@tcp1  failover.node=192.168.227.12@tcp1:192.168.227.11@tcp1

checking for existing Lustre data: not found
mkfs_cmd = zfs create -o canmount=off -o xattr=sa demo-mdt0pool/mdt0
Writing demo-mdt0pool/mdt0 properties
  lustre:version=1
  lustre:flags=4193
  lustre:index=0
  lustre:fsname=demo
  lustre:svname=demo:MDT0000
  lustre:mgsnode=192.168.227.11@tcp1:192.168.227.12@tcp1
  lustre:failover.node=192.168.227.12@tcp1:192.168.227.11@tcp1

Use the zfs get command or tunefs.lustre to verify that the file system dataset has been formatted correctly. For example:

[root@rh7z-mds2 ~]# zfs get all -s local
NAME                PROPERTY              VALUE                                    SOURCE
demo-mdt0pool       canmount              off                                      local
demo-mdt0pool/mdt0  canmount              off                                      local
demo-mdt0pool/mdt0  xattr                 sa                                       local
demo-mdt0pool/mdt0  lustre:svname         demo-MDT0000                             local
demo-mdt0pool/mdt0  lustre:flags          4129                                     local
demo-mdt0pool/mdt0  lustre:failover.node  192.168.227.12@tcp1:192.168.227.11@tcp1  local
demo-mdt0pool/mdt0  lustre:version        1                                        local
demo-mdt0pool/mdt0  lustre:mgsnode        192.168.227.11@tcp1:192.168.227.12@tcp1  local
demo-mdt0pool/mdt0  lustre:fsname         demo                                     local
demo-mdt0pool/mdt0  lustre:index          0                                        local

Starting the MDS Service

The mount command is used to start all Lustre storage services, including the MDS. The syntax is:

mount -t lustre [-o <options>] \
  <ldiskfs blockdev>|<zpool>/<dataset> <mount point>

The mount command syntax is very similar for both LDISKFS and ZFS storage targets. The main difference is the format of the path to the storage. For LDISKFS, the path will resolve to a block device, such as /dev/sda or /dev/mapper/mpatha, whereas for ZFS, the path resolves to a dataset in a zpool, e.g. demo-mdt0pool/mdt0.

The mount point directory must exist before the mount command is executed. The recommended convention for the mount point of the MDT storage is /lustre/<fsname>/mdt<n>, where <fsname> is the name of the file system and <n> is the index number of the MDT.

The following example starts a ZFS-based MDS:

# Ignore MOUNTPOINT column in output: not used by Lustre
[root@rh7z-mds2 ~]# zfs list
NAME                 USED  AVAIL  REFER  MOUNTPOINT
demo-mdt0pool       2.87M  9.62G    19K  /demo-mdt0pool
demo-mdt0pool/mdt0  2.79M  9.62G  2.79M  /demo-mdt0pool/mdt0

[root@rh7z-mds2 ~]# mkdir -p /lustre/demo/mdt0

[root@rh7z-mds2 ~]# mount -t lustre demo-mdt0pool/mdt0 /lustre/demo/mdt0

[root@rh7z-mds2 ~]# df -ht lustre
File system          Size  Used Avail Use% Mounted on
demo-mdt0pool/mdt0  9.7G  2.8M  9.7G   1% /lustre/demo/mdt0

Note: The default output for zfs list shows mount points for the demo-mdt0pool Pool and mdt0 dataset in the MOUNTPOINT column. Just as for all ZFS-formatted OSDs, the content in this column can be ignored.

To reduce confusion, the ZFS file system mountpoint property can be set equal to none. For example:

zfs set mountpoint=none demo-mdt0pool
zfs set mountpoint=none demo-mdt0pool/mdt0

Note: Only the mount -t lustre command can start Lustre services. Mounting storage as type ldiskfs or zfs will mount a storage target on the host, but it will not trigger the startup of the requisite Lustre kernel processes.

To verify that the MDS is running, check that the MDT device has been mounted, then get the Lustre device list with lctl dl, and review the running processes:

[root@rh7z-mds2 ~]# df -ht lustre
File system          Size  Used Avail Use% Mounted on
demo-mdt0pool/mdt0  9.7G  2.8M  9.7G   1% /lustre/demo/mdt0

[root@rh7z-mds2 ~]# lctl dl
  0 UP osd-zfs demo-MDT0000-osd demo-MDT0000-osd_UUID 7
  1 UP mgc MGC192.168.227.11@tcp1 1605562b-d702-9251-6f38-1fd4a64e2720 5
  2 UP mds MDS MDS_uuid 3
  3 UP lod demo-MDT0000-mdtlov demo-MDT0000-mdtlov_UUID 4
  4 UP mdt demo-MDT0000 demo-MDT0000_UUID 5
  5 UP mdd demo-MDD0000 demo-MDD0000_UUID 4
  6 UP qmt demo-QMT0000 demo-QMT0000_UUID 4
  7 UP lwp demo-MDT0000-lwp-MDT0000 demo-MDT0000-lwp-MDT0000_UUID 5

[root@rh7z-mds2 ~]# ps -ef | awk '/mdt/ && !/awk/'
root     32320     2  0 Mar30 ?        00:00:00 [mdt00_000]
root     32321     2  0 Mar30 ?        00:00:00 [mdt00_001]
root     32322     2  0 Mar30 ?        00:00:00 [mdt00_002]
root     32323     2  0 Mar30 ?        00:00:00 [mdt_rdpg00_000]
root     32324     2  0 Mar30 ?        00:00:00 [mdt_rdpg00_001]
root     32325     2  0 Mar30 ?        00:00:00 [mdt_attr00_000]
root     32326     2  0 Mar30 ?        00:00:00 [mdt_attr00_001]
root     32327     2  0 Mar30 ?        00:00:00 [mdt_out00_000]
root     32328     2  0 Mar30 ?        00:00:00 [mdt_out00_001]
root     32329     2  0 Mar30 ?        00:00:00 [mdt_seqs_0000]
root     32330     2  0 Mar30 ?        00:00:00 [mdt_seqs_0001]
root     32331     2  0 Mar30 ?        00:00:00 [mdt_seqm_0000]
root     32332     2  0 Mar30 ?        00:00:00 [mdt_seqm_0001]
root     32333     2  0 Mar30 ?        00:00:00 [mdt_fld_0000]
root     32334     2  0 Mar30 ?        00:00:00 [mdt_fld_0001]
root     32340     2  0 Mar30 ?        00:00:00 [mdt_ck]

Stopping the MDS Service

To stop a Lustre service, umount the corresponding target:

umount <mount point>

The mount point must correspond to the mount point used with the mount -t lustre command. For example:

[root@rh7z-mds2 ~]# df -ht lustre
File system          Size  Used Avail Use% Mounted on
demo-mdt0pool/mdt0  9.7G  2.8M  9.7G   1% /lustre/demo/mdt0
[root@rh7z-mds2 ~]# umount /lustre/demo/mdt0
[root@rh7z-mds2 ~]# df -ht lustre
df: no file systems processed
[root@rh7z-mds2 ~]# lctl dl
[root@rh7z-mds2 ~]#

Using the regular umount command is the correct way to stop a given Lustre service and unmount the associated storage, for both LDISKFS and ZFS-based Lustre storage volumes.

Do not use the zfs unmount command to stop a Lustre service. Attempting to use zfs commands to unmount a storage target that is mounted as part of an active Lustre service will return an error.